Batch - how to choose machine type?

[CloudSim]

Hi,

I'm a little confused about how select InstancePolicy.machine_type when creating jobs through Batch.

Let's say that I have a task which needs 2000 milliCpu and 2000 MiB and I want to run 200 of them. What should I choose for a machine type? If I want complete parallelism (e.g. 200) does that change anything?

Some particular cases I don't understand from the API documentation:

• If I choose n1-standard-1 with only one vCPU will it fail to run any tasks?
• If I choose n1-standard-4 will it run 2 tasks per VM? Will it be wasting 15 - 2*2 = 11 GB of ram?
• Can I leave machine_type unspecified and have it automatically determined for efficiency?

Thank you!

[bolianyin]

Batch has a goal to choose machine type intelligently if users don't specify one, but there is a lot room for improvement today. In the particular examples, I think the current behaviors are

. If I choose n1-standard-1 with only one vCPU will it fail to run any tasks?

>> Batch will try to schedule 1 task per VM trusting that is the intention of the user.

. If I choose n1-standard-4 will it run 2 tasks per VM? Will it be wasting 15 - 2*2 = 11 GB of ram?

>> Yes. Batch uses CPU numbers to infer one VM can run at most 2 tasks.

. Can I leave machine_type unspecified and have it automatically determined for efficiency?

>> Yes. Today Batch will select an e2 machine although the behavior could change.

[CloudSim]

Thank you. I will experiment with leaving the machine type unspecified.

[CloudSim]

Hi,

That worked well for me, thank you! Just a brief follow up question:

What if I don't specify CPU requirements in compute resources or machine type in an instance policy? Can Batch autoscale my job so that each task is automatically provisioned with sufficient CPU?

Thanks

[bolianyin]

In that case, Batch will apply a pretty low default CPU requirements to tasks (which may not be sufficient) and use that to come up with a machine type.